net/ipv4/inet_hashtables.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   4  *              operating system.  INET is implemented using the BSD Socket
   5  *              interface as the means of communication with the user level.
   6  *
   7  *              Generic INET transport hashtables
   8  *
   9  * Authors:     Lotsa people, from code originally in tcp
  10  */
  11
  12 #include <linux/module.h>
  13 #include <linux/random.h>
  14 #include <linux/sched.h>
  15 #include <linux/slab.h>
  16 #include <linux/wait.h>
  17 #include <linux/vmalloc.h>
  18 #include <linux/memblock.h>
  19
  20 #include <net/addrconf.h>
  21 #include <net/inet_connection_sock.h>
  22 #include <net/inet_hashtables.h>
  23 #if IS_ENABLED(CONFIG_IPV6)
  24 #include <net/inet6_hashtables.h>
  25 #endif
  26 #include <net/secure_seq.h>
  27 #include <net/ip.h>
  28 #include <net/tcp.h>
  29 #include <net/sock_reuseport.h>
  30
  31 static u32 inet_ehashfn(const struct net *net, const __be32 laddr,
  32                         const __u16 lport, const __be32 faddr,
  33                         const __be16 fport)
  34 {
  35         static u32 inet_ehash_secret __read_mostly;
  36
  37         net_get_random_once(&inet_ehash_secret, sizeof(inet_ehash_secret));
  38
  39         return __inet_ehashfn(laddr, lport, faddr, fport,
  40                               inet_ehash_secret + net_hash_mix(net));
  41 }
  42
  43 /* This function handles inet_sock, but also timewait and request sockets
  44  * for IPv4/IPv6.
  45  */
  46 static u32 sk_ehashfn(const struct sock *sk)
  47 {
  48 #if IS_ENABLED(CONFIG_IPV6)
  49         if (sk->sk_family == AF_INET6 &&
  50             !ipv6_addr_v4mapped(&sk->sk_v6_daddr))
  51                 return inet6_ehashfn(sock_net(sk),
  52                                      &sk->sk_v6_rcv_saddr, sk->sk_num,
  53                                      &sk->sk_v6_daddr, sk->sk_dport);
  54 #endif
  55         return inet_ehashfn(sock_net(sk),
  56                             sk->sk_rcv_saddr, sk->sk_num,
  57                             sk->sk_daddr, sk->sk_dport);
  58 }
  59
  60 /*
  61  * Allocate and initialize a new local port bind bucket.
  62  * The bindhash mutex for snum's hash chain must be held here.
  63  */
  64 struct inet_bind_bucket *inet_bind_bucket_create(struct kmem_cache *cachep,
  65                                                  struct net *net,
  66                                                  struct inet_bind_hashbucket *head,
  67                                                  const unsigned short snum,
  68                                                  int l3mdev)
  69 {
  70         struct inet_bind_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
  71
  72         if (tb) {
  73                 write_pnet(&tb->ib_net, net);
  74                 tb->l3mdev    = l3mdev;
  75                 tb->port      = snum;
  76                 tb->fastreuse = 0;
  77                 tb->fastreuseport = 0;
  78                 INIT_HLIST_HEAD(&tb->owners);
  79                 hlist_add_head(&tb->node, &head->chain);
  80         }
  81         return tb;
  82 }
  83
  84 /*
  85  * Caller must hold hashbucket lock for this tb with local BH disabled
  86  */
  87 void inet_bind_bucket_destroy(struct kmem_cache *cachep, struct inet_bind_bucket *tb)
  88 {
  89         if (hlist_empty(&tb->owners)) {
  90                 __hlist_del(&tb->node);
  91                 kmem_cache_free(cachep, tb);
  92         }
  93 }
  94
  95 bool inet_bind_bucket_match(const struct inet_bind_bucket *tb, const struct net *net,
  96                             unsigned short port, int l3mdev)
  97 {
  98         return net_eq(ib_net(tb), net) && tb->port == port &&
  99                 tb->l3mdev == l3mdev;
 100 }
 101
 102 static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb,
 103                                    struct net *net,
 104                                    struct inet_bind_hashbucket *head,
 105                                    unsigned short port, int l3mdev,
 106                                    const struct sock *sk)
 107 {
 108         write_pnet(&tb->ib_net, net);
 109         tb->l3mdev    = l3mdev;
 110         tb->port      = port;
 111 #if IS_ENABLED(CONFIG_IPV6)
 112         if (sk->sk_family == AF_INET6)
 113                 tb->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
 114         else
 115 #endif
 116                 tb->rcv_saddr = sk->sk_rcv_saddr;
 117         INIT_HLIST_HEAD(&tb->owners);
 118         hlist_add_head(&tb->node, &head->chain);
 119 }
 120
 121 struct inet_bind2_bucket *inet_bind2_bucket_create(struct kmem_cache *cachep,
 122                                                    struct net *net,
 123                                                    struct inet_bind_hashbucket *head,
 124                                                    unsigned short port,
 125                                                    int l3mdev,
 126                                                    const struct sock *sk)
 127 {
 128         struct inet_bind2_bucket *tb = kmem_cache_alloc(cachep, GFP_ATOMIC);
 129
 130         if (tb)
 131                 inet_bind2_bucket_init(tb, net, head, port, l3mdev, sk);
 132
 133         return tb;
 134 }
 135
 136 /* Caller must hold hashbucket lock for this tb with local BH disabled */
 137 void inet_bind2_bucket_destroy(struct kmem_cache *cachep, struct inet_bind2_bucket *tb)
 138 {
 139         if (hlist_empty(&tb->owners)) {
 140                 __hlist_del(&tb->node);
 141                 kmem_cache_free(cachep, tb);
 142         }
 143 }
 144
 145 static bool inet_bind2_bucket_addr_match(const struct inet_bind2_bucket *tb2,
 146                                          const struct sock *sk)
 147 {
 148 #if IS_ENABLED(CONFIG_IPV6)
 149         if (sk->sk_family == AF_INET6)
 150                 return ipv6_addr_equal(&tb2->v6_rcv_saddr,
 151                                        &sk->sk_v6_rcv_saddr);
 152 #endif
 153         return tb2->rcv_saddr == sk->sk_rcv_saddr;
 154 }
 155
 156 void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,
 157                     struct inet_bind2_bucket *tb2, unsigned short port)
 158 {
 159         inet_sk(sk)->inet_num = port;
 160         sk_add_bind_node(sk, &tb->owners);
 161         inet_csk(sk)->icsk_bind_hash = tb;
 162         sk_add_bind2_node(sk, &tb2->owners);
 163         inet_csk(sk)->icsk_bind2_hash = tb2;
 164 }
 165
 166 /*
 167  * Get rid of any references to a local port held by the given sock.
 168  */
 169 static void __inet_put_port(struct sock *sk)
 170 {
 171         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 172         struct inet_bind_hashbucket *head, *head2;
 173         struct net *net = sock_net(sk);
 174         struct inet_bind_bucket *tb;
 175         int bhash;
 176
 177         bhash = inet_bhashfn(net, inet_sk(sk)->inet_num, hashinfo->bhash_size);
 178         head = &hashinfo->bhash[bhash];
 179         head2 = inet_bhashfn_portaddr(hashinfo, sk, net, inet_sk(sk)->inet_num);
 180
 181         spin_lock(&head->lock);
 182         tb = inet_csk(sk)->icsk_bind_hash;
 183         __sk_del_bind_node(sk);
 184         inet_csk(sk)->icsk_bind_hash = NULL;
 185         inet_sk(sk)->inet_num = 0;
 186         inet_bind_bucket_destroy(hashinfo->bind_bucket_cachep, tb);
 187
 188         spin_lock(&head2->lock);
 189         if (inet_csk(sk)->icsk_bind2_hash) {
 190                 struct inet_bind2_bucket *tb2 = inet_csk(sk)->icsk_bind2_hash;
 191
 192                 __sk_del_bind2_node(sk);
 193                 inet_csk(sk)->icsk_bind2_hash = NULL;
 194                 inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
 195         }
 196         spin_unlock(&head2->lock);
 197
 198         spin_unlock(&head->lock);
 199 }
 200
 201 void inet_put_port(struct sock *sk)
 202 {
 203         local_bh_disable();
 204         __inet_put_port(sk);
 205         local_bh_enable();
 206 }
 207 EXPORT_SYMBOL(inet_put_port);
 208
 209 int __inet_inherit_port(const struct sock *sk, struct sock *child)
 210 {
 211         struct inet_hashinfo *table = tcp_or_dccp_get_hashinfo(sk);
 212         unsigned short port = inet_sk(child)->inet_num;
 213         struct inet_bind_hashbucket *head, *head2;
 214         bool created_inet_bind_bucket = false;
 215         struct net *net = sock_net(sk);
 216         bool update_fastreuse = false;
 217         struct inet_bind2_bucket *tb2;
 218         struct inet_bind_bucket *tb;
 219         int bhash, l3mdev;
 220
 221         bhash = inet_bhashfn(net, port, table->bhash_size);
 222         head = &table->bhash[bhash];
 223         head2 = inet_bhashfn_portaddr(table, child, net, port);
 224
 225         spin_lock(&head->lock);
 226         spin_lock(&head2->lock);
 227         tb = inet_csk(sk)->icsk_bind_hash;
 228         tb2 = inet_csk(sk)->icsk_bind2_hash;
 229         if (unlikely(!tb || !tb2)) {
 230                 spin_unlock(&head2->lock);
 231                 spin_unlock(&head->lock);
 232                 return -ENOENT;
 233         }
 234         if (tb->port != port) {
 235                 l3mdev = inet_sk_bound_l3mdev(sk);
 236
 237                 /* NOTE: using tproxy and redirecting skbs to a proxy
 238                  * on a different listener port breaks the assumption
 239                  * that the listener socket's icsk_bind_hash is the same
 240                  * as that of the child socket. We have to look up or
 241                  * create a new bind bucket for the child here. */
 242                 inet_bind_bucket_for_each(tb, &head->chain) {
 243                         if (inet_bind_bucket_match(tb, net, port, l3mdev))
 244                                 break;
 245                 }
 246                 if (!tb) {
 247                         tb = inet_bind_bucket_create(table->bind_bucket_cachep,
 248                                                      net, head, port, l3mdev);
 249                         if (!tb) {
 250                                 spin_unlock(&head2->lock);
 251                                 spin_unlock(&head->lock);
 252                                 return -ENOMEM;
 253                         }
 254                         created_inet_bind_bucket = true;
 255                 }
 256                 update_fastreuse = true;
 257
 258                 goto bhash2_find;
 259         } else if (!inet_bind2_bucket_addr_match(tb2, child)) {
 260                 l3mdev = inet_sk_bound_l3mdev(sk);
 261
 262 bhash2_find:
 263                 tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, child);
 264                 if (!tb2) {
 265                         tb2 = inet_bind2_bucket_create(table->bind2_bucket_cachep,
 266                                                        net, head2, port,
 267                                                        l3mdev, child);
 268                         if (!tb2)
 269                                 goto error;
 270                 }
 271         }
 272         if (update_fastreuse)
 273                 inet_csk_update_fastreuse(tb, child);
 274         inet_bind_hash(child, tb, tb2, port);
 275         spin_unlock(&head2->lock);
 276         spin_unlock(&head->lock);
 277
 278         return 0;
 279
 280 error:
 281         if (created_inet_bind_bucket)
 282                 inet_bind_bucket_destroy(table->bind_bucket_cachep, tb);
 283         spin_unlock(&head2->lock);
 284         spin_unlock(&head->lock);
 285         return -ENOMEM;
 286 }
 287 EXPORT_SYMBOL_GPL(__inet_inherit_port);
 288
 289 static struct inet_listen_hashbucket *
 290 inet_lhash2_bucket_sk(struct inet_hashinfo *h, struct sock *sk)
 291 {
 292         u32 hash;
 293
 294 #if IS_ENABLED(CONFIG_IPV6)
 295         if (sk->sk_family == AF_INET6)
 296                 hash = ipv6_portaddr_hash(sock_net(sk),
 297                                           &sk->sk_v6_rcv_saddr,
 298                                           inet_sk(sk)->inet_num);
 299         else
 300 #endif
 301                 hash = ipv4_portaddr_hash(sock_net(sk),
 302                                           inet_sk(sk)->inet_rcv_saddr,
 303                                           inet_sk(sk)->inet_num);
 304         return inet_lhash2_bucket(h, hash);
 305 }
 306
 307 static inline int compute_score(struct sock *sk, struct net *net,
 308                                 const unsigned short hnum, const __be32 daddr,
 309                                 const int dif, const int sdif)
 310 {
 311         int score = -1;
 312
 313         if (net_eq(sock_net(sk), net) && sk->sk_num == hnum &&
 314                         !ipv6_only_sock(sk)) {
 315                 if (sk->sk_rcv_saddr != daddr)
 316                         return -1;
 317
 318                 if (!inet_sk_bound_dev_eq(net, sk->sk_bound_dev_if, dif, sdif))
 319                         return -1;
 320                 score =  sk->sk_bound_dev_if ? 2 : 1;
 321
 322                 if (sk->sk_family == PF_INET)
 323                         score++;
 324                 if (READ_ONCE(sk->sk_incoming_cpu) == raw_smp_processor_id())
 325                         score++;
 326         }
 327         return score;
 328 }
 329
 330 static inline struct sock *lookup_reuseport(struct net *net, struct sock *sk,
 331                                             struct sk_buff *skb, int doff,
 332                                             __be32 saddr, __be16 sport,
 333                                             __be32 daddr, unsigned short hnum)
 334 {
 335         struct sock *reuse_sk = NULL;
 336         u32 phash;
 337
 338         if (sk->sk_reuseport) {
 339                 phash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 340                 reuse_sk = reuseport_select_sock(sk, phash, skb, doff);
 341         }
 342         return reuse_sk;
 343 }
 344
 345 /*
 346  * Here are some nice properties to exploit here. The BSD API
 347  * does not allow a listening sock to specify the remote port nor the
 348  * remote address for the connection. So always assume those are both
 349  * wildcarded during the search since they can never be otherwise.
 350  */
 351
 352 /* called with rcu_read_lock() : No refcount taken on the socket */
 353 static struct sock *inet_lhash2_lookup(struct net *net,
 354                                 struct inet_listen_hashbucket *ilb2,
 355                                 struct sk_buff *skb, int doff,
 356                                 const __be32 saddr, __be16 sport,
 357                                 const __be32 daddr, const unsigned short hnum,
 358                                 const int dif, const int sdif)
 359 {
 360         struct sock *sk, *result = NULL;
 361         struct hlist_nulls_node *node;
 362         int score, hiscore = 0;
 363
 364         sk_nulls_for_each_rcu(sk, node, &ilb2->nulls_head) {
 365                 score = compute_score(sk, net, hnum, daddr, dif, sdif);
 366                 if (score > hiscore) {
 367                         result = lookup_reuseport(net, sk, skb, doff,
 368                                                   saddr, sport, daddr, hnum);
 369                         if (result)
 370                                 return result;
 371
 372                         result = sk;
 373                         hiscore = score;
 374                 }
 375         }
 376
 377         return result;
 378 }
 379
 380 static inline struct sock *inet_lookup_run_bpf(struct net *net,
 381                                                struct inet_hashinfo *hashinfo,
 382                                                struct sk_buff *skb, int doff,
 383                                                __be32 saddr, __be16 sport,
 384                                                __be32 daddr, u16 hnum, const int dif)
 385 {
 386         struct sock *sk, *reuse_sk;
 387         bool no_reuseport;
 388
 389         if (hashinfo != net->ipv4.tcp_death_row.hashinfo)
 390                 return NULL; /* only TCP is supported */
 391
 392         no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP, saddr, sport,
 393                                             daddr, hnum, dif, &sk);
 394         if (no_reuseport || IS_ERR_OR_NULL(sk))
 395                 return sk;
 396
 397         reuse_sk = lookup_reuseport(net, sk, skb, doff, saddr, sport, daddr, hnum);
 398         if (reuse_sk)
 399                 sk = reuse_sk;
 400         return sk;
 401 }
 402
 403 struct sock *__inet_lookup_listener(struct net *net,
 404                                     struct inet_hashinfo *hashinfo,
 405                                     struct sk_buff *skb, int doff,
 406                                     const __be32 saddr, __be16 sport,
 407                                     const __be32 daddr, const unsigned short hnum,
 408                                     const int dif, const int sdif)
 409 {
 410         struct inet_listen_hashbucket *ilb2;
 411         struct sock *result = NULL;
 412         unsigned int hash2;
 413
 414         /* Lookup redirect from BPF */
 415         if (static_branch_unlikely(&bpf_sk_lookup_enabled)) {
 416                 result = inet_lookup_run_bpf(net, hashinfo, skb, doff,
 417                                              saddr, sport, daddr, hnum, dif);
 418                 if (result)
 419                         goto done;
 420         }
 421
 422         hash2 = ipv4_portaddr_hash(net, daddr, hnum);
 423         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 424
 425         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 426                                     saddr, sport, daddr, hnum,
 427                                     dif, sdif);
 428         if (result)
 429                 goto done;
 430
 431         /* Lookup lhash2 with INADDR_ANY */
 432         hash2 = ipv4_portaddr_hash(net, htonl(INADDR_ANY), hnum);
 433         ilb2 = inet_lhash2_bucket(hashinfo, hash2);
 434
 435         result = inet_lhash2_lookup(net, ilb2, skb, doff,
 436                                     saddr, sport, htonl(INADDR_ANY), hnum,
 437                                     dif, sdif);
 438 done:
 439         if (IS_ERR(result))
 440                 return NULL;
 441         return result;
 442 }
 443 EXPORT_SYMBOL_GPL(__inet_lookup_listener);
 444
 445 /* All sockets share common refcount, but have different destructors */
 446 void sock_gen_put(struct sock *sk)
 447 {
 448         if (!refcount_dec_and_test(&sk->sk_refcnt))
 449                 return;
 450
 451         if (sk->sk_state == TCP_TIME_WAIT)
 452                 inet_twsk_free(inet_twsk(sk));
 453         else if (sk->sk_state == TCP_NEW_SYN_RECV)
 454                 reqsk_free(inet_reqsk(sk));
 455         else
 456                 sk_free(sk);
 457 }
 458 EXPORT_SYMBOL_GPL(sock_gen_put);
 459
 460 void sock_edemux(struct sk_buff *skb)
 461 {
 462         sock_gen_put(skb->sk);
 463 }
 464 EXPORT_SYMBOL(sock_edemux);
 465
 466 struct sock *__inet_lookup_established(struct net *net,
 467                                   struct inet_hashinfo *hashinfo,
 468                                   const __be32 saddr, const __be16 sport,
 469                                   const __be32 daddr, const u16 hnum,
 470                                   const int dif, const int sdif)
 471 {
 472         INET_ADDR_COOKIE(acookie, saddr, daddr);
 473         const __portpair ports = INET_COMBINED_PORTS(sport, hnum);
 474         struct sock *sk;
 475         const struct hlist_nulls_node *node;
 476         /* Optimize here for direct hit, only listening connections can
 477          * have wildcards anyways.
 478          */
 479         unsigned int hash = inet_ehashfn(net, daddr, hnum, saddr, sport);
 480         unsigned int slot = hash & hashinfo->ehash_mask;
 481         struct inet_ehash_bucket *head = &hashinfo->ehash[slot];
 482
 483 begin:
 484         sk_nulls_for_each_rcu(sk, node, &head->chain) {
 485                 if (sk->sk_hash != hash)
 486                         continue;
 487                 if (likely(inet_match(net, sk, acookie, ports, dif, sdif))) {
 488                         if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
 489                                 goto out;
 490                         if (unlikely(!inet_match(net, sk, acookie,
 491                                                  ports, dif, sdif))) {
 492                                 sock_gen_put(sk);
 493                                 goto begin;
 494                         }
 495                         goto found;
 496                 }
 497         }
 498         /*
 499          * if the nulls value we got at the end of this lookup is
 500          * not the expected one, we must restart lookup.
 501          * We probably met an item that was moved to another chain.
 502          */
 503         if (get_nulls_value(node) != slot)
 504                 goto begin;
 505 out:
 506         sk = NULL;
 507 found:
 508         return sk;
 509 }
 510 EXPORT_SYMBOL_GPL(__inet_lookup_established);
 511
 512 /* called with local bh disabled */
 513 static int __inet_check_established(struct inet_timewait_death_row *death_row,
 514                                     struct sock *sk, __u16 lport,
 515                                     struct inet_timewait_sock **twp)
 516 {
 517         struct inet_hashinfo *hinfo = death_row->hashinfo;
 518         struct inet_sock *inet = inet_sk(sk);
 519         __be32 daddr = inet->inet_rcv_saddr;
 520         __be32 saddr = inet->inet_daddr;
 521         int dif = sk->sk_bound_dev_if;
 522         struct net *net = sock_net(sk);
 523         int sdif = l3mdev_master_ifindex_by_index(net, dif);
 524         INET_ADDR_COOKIE(acookie, saddr, daddr);
 525         const __portpair ports = INET_COMBINED_PORTS(inet->inet_dport, lport);
 526         unsigned int hash = inet_ehashfn(net, daddr, lport,
 527                                          saddr, inet->inet_dport);
 528         struct inet_ehash_bucket *head = inet_ehash_bucket(hinfo, hash);
 529         spinlock_t *lock = inet_ehash_lockp(hinfo, hash);
 530         struct sock *sk2;
 531         const struct hlist_nulls_node *node;
 532         struct inet_timewait_sock *tw = NULL;
 533
 534         spin_lock(lock);
 535
 536         sk_nulls_for_each(sk2, node, &head->chain) {
 537                 if (sk2->sk_hash != hash)
 538                         continue;
 539
 540                 if (likely(inet_match(net, sk2, acookie, ports, dif, sdif))) {
 541                         if (sk2->sk_state == TCP_TIME_WAIT) {
 542                                 tw = inet_twsk(sk2);
 543                                 if (twsk_unique(sk, sk2, twp))
 544                                         break;
 545                         }
 546                         goto not_unique;
 547                 }
 548         }
 549
 550         /* Must record num and sport now. Otherwise we will see
 551          * in hash table socket with a funny identity.
 552          */
 553         inet->inet_num = lport;
 554         inet->inet_sport = htons(lport);
 555         sk->sk_hash = hash;
 556         WARN_ON(!sk_unhashed(sk));
 557         __sk_nulls_add_node_rcu(sk, &head->chain);
 558         if (tw) {
 559                 sk_nulls_del_node_init_rcu((struct sock *)tw);
 560                 __NET_INC_STATS(net, LINUX_MIB_TIMEWAITRECYCLED);
 561         }
 562         spin_unlock(lock);
 563         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 564
 565         if (twp) {
 566                 *twp = tw;
 567         } else if (tw) {
 568                 /* Silly. Should hash-dance instead... */
 569                 inet_twsk_deschedule_put(tw);
 570         }
 571         return 0;
 572
 573 not_unique:
 574         spin_unlock(lock);
 575         return -EADDRNOTAVAIL;
 576 }
 577
 578 static u64 inet_sk_port_offset(const struct sock *sk)
 579 {
 580         const struct inet_sock *inet = inet_sk(sk);
 581
 582         return secure_ipv4_port_ephemeral(inet->inet_rcv_saddr,
 583                                           inet->inet_daddr,
 584                                           inet->inet_dport);
 585 }
 586
 587 /* Searches for an exsiting socket in the ehash bucket list.
 588  * Returns true if found, false otherwise.
 589  */
 590 static bool inet_ehash_lookup_by_sk(struct sock *sk,
 591                                     struct hlist_nulls_head *list)
 592 {
 593         const __portpair ports = INET_COMBINED_PORTS(sk->sk_dport, sk->sk_num);
 594         const int sdif = sk->sk_bound_dev_if;
 595         const int dif = sk->sk_bound_dev_if;
 596         const struct hlist_nulls_node *node;
 597         struct net *net = sock_net(sk);
 598         struct sock *esk;
 599
 600         INET_ADDR_COOKIE(acookie, sk->sk_daddr, sk->sk_rcv_saddr);
 601
 602         sk_nulls_for_each_rcu(esk, node, list) {
 603                 if (esk->sk_hash != sk->sk_hash)
 604                         continue;
 605                 if (sk->sk_family == AF_INET) {
 606                         if (unlikely(inet_match(net, esk, acookie,
 607                                                 ports, dif, sdif))) {
 608                                 return true;
 609                         }
 610                 }
 611 #if IS_ENABLED(CONFIG_IPV6)
 612                 else if (sk->sk_family == AF_INET6) {
 613                         if (unlikely(inet6_match(net, esk,
 614                                                  &sk->sk_v6_daddr,
 615                                                  &sk->sk_v6_rcv_saddr,
 616                                                  ports, dif, sdif))) {
 617                                 return true;
 618                         }
 619                 }
 620 #endif
 621         }
 622         return false;
 623 }
 624
 625 /* Insert a socket into ehash, and eventually remove another one
 626  * (The another one can be a SYN_RECV or TIMEWAIT)
 627  * If an existing socket already exists, socket sk is not inserted,
 628  * and sets found_dup_sk parameter to true.
 629  */
 630 bool inet_ehash_insert(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 631 {
 632         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 633         struct inet_ehash_bucket *head;
 634         struct hlist_nulls_head *list;
 635         spinlock_t *lock;
 636         bool ret = true;
 637
 638         WARN_ON_ONCE(!sk_unhashed(sk));
 639
 640         sk->sk_hash = sk_ehashfn(sk);
 641         head = inet_ehash_bucket(hashinfo, sk->sk_hash);
 642         list = &head->chain;
 643         lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 644
 645         spin_lock(lock);
 646         if (osk) {
 647                 WARN_ON_ONCE(sk->sk_hash != osk->sk_hash);
 648                 ret = sk_nulls_del_node_init_rcu(osk);
 649         } else if (found_dup_sk) {
 650                 *found_dup_sk = inet_ehash_lookup_by_sk(sk, list);
 651                 if (*found_dup_sk)
 652                         ret = false;
 653         }
 654
 655         if (ret)
 656                 __sk_nulls_add_node_rcu(sk, list);
 657
 658         spin_unlock(lock);
 659
 660         return ret;
 661 }
 662
 663 bool inet_ehash_nolisten(struct sock *sk, struct sock *osk, bool *found_dup_sk)
 664 {
 665         bool ok = inet_ehash_insert(sk, osk, found_dup_sk);
 666
 667         if (ok) {
 668                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 669         } else {
 670                 this_cpu_inc(*sk->sk_prot->orphan_count);
 671                 inet_sk_set_state(sk, TCP_CLOSE);
 672                 sock_set_flag(sk, SOCK_DEAD);
 673                 inet_csk_destroy_sock(sk);
 674         }
 675         return ok;
 676 }
 677 EXPORT_SYMBOL_GPL(inet_ehash_nolisten);
 678
 679 static int inet_reuseport_add_sock(struct sock *sk,
 680                                    struct inet_listen_hashbucket *ilb)
 681 {
 682         struct inet_bind_bucket *tb = inet_csk(sk)->icsk_bind_hash;
 683         const struct hlist_nulls_node *node;
 684         struct sock *sk2;
 685         kuid_t uid = sock_i_uid(sk);
 686
 687         sk_nulls_for_each_rcu(sk2, node, &ilb->nulls_head) {
 688                 if (sk2 != sk &&
 689                     sk2->sk_family == sk->sk_family &&
 690                     ipv6_only_sock(sk2) == ipv6_only_sock(sk) &&
 691                     sk2->sk_bound_dev_if == sk->sk_bound_dev_if &&
 692                     inet_csk(sk2)->icsk_bind_hash == tb &&
 693                     sk2->sk_reuseport && uid_eq(uid, sock_i_uid(sk2)) &&
 694                     inet_rcv_saddr_equal(sk, sk2, false))
 695                         return reuseport_add_sock(sk, sk2,
 696                                                   inet_rcv_saddr_any(sk));
 697         }
 698
 699         return reuseport_alloc(sk, inet_rcv_saddr_any(sk));
 700 }
 701
 702 int __inet_hash(struct sock *sk, struct sock *osk)
 703 {
 704         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 705         struct inet_listen_hashbucket *ilb2;
 706         int err = 0;
 707
 708         if (sk->sk_state != TCP_LISTEN) {
 709                 local_bh_disable();
 710                 inet_ehash_nolisten(sk, osk, NULL);
 711                 local_bh_enable();
 712                 return 0;
 713         }
 714         WARN_ON(!sk_unhashed(sk));
 715         ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 716
 717         spin_lock(&ilb2->lock);
 718         if (sk->sk_reuseport) {
 719                 err = inet_reuseport_add_sock(sk, ilb2);
 720                 if (err)
 721                         goto unlock;
 722         }
 723         if (IS_ENABLED(CONFIG_IPV6) && sk->sk_reuseport &&
 724                 sk->sk_family == AF_INET6)
 725                 __sk_nulls_add_node_tail_rcu(sk, &ilb2->nulls_head);
 726         else
 727                 __sk_nulls_add_node_rcu(sk, &ilb2->nulls_head);
 728         sock_set_flag(sk, SOCK_RCU_FREE);
 729         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 730 unlock:
 731         spin_unlock(&ilb2->lock);
 732
 733         return err;
 734 }
 735 EXPORT_SYMBOL(__inet_hash);
 736
 737 int inet_hash(struct sock *sk)
 738 {
 739         int err = 0;
 740
 741         if (sk->sk_state != TCP_CLOSE)
 742                 err = __inet_hash(sk, NULL);
 743
 744         return err;
 745 }
 746 EXPORT_SYMBOL_GPL(inet_hash);
 747
 748 void inet_unhash(struct sock *sk)
 749 {
 750         struct inet_hashinfo *hashinfo = tcp_or_dccp_get_hashinfo(sk);
 751
 752         if (sk_unhashed(sk))
 753                 return;
 754
 755         if (sk->sk_state == TCP_LISTEN) {
 756                 struct inet_listen_hashbucket *ilb2;
 757
 758                 ilb2 = inet_lhash2_bucket_sk(hashinfo, sk);
 759                 /* Don't disable bottom halves while acquiring the lock to
 760                  * avoid circular locking dependency on PREEMPT_RT.
 761                  */
 762                 spin_lock(&ilb2->lock);
 763                 if (sk_unhashed(sk)) {
 764                         spin_unlock(&ilb2->lock);
 765                         return;
 766                 }
 767
 768                 if (rcu_access_pointer(sk->sk_reuseport_cb))
 769                         reuseport_stop_listen_sock(sk);
 770
 771                 __sk_nulls_del_node_init_rcu(sk);
 772                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 773                 spin_unlock(&ilb2->lock);
 774         } else {
 775                 spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
 776
 777                 spin_lock_bh(lock);
 778                 if (sk_unhashed(sk)) {
 779                         spin_unlock_bh(lock);
 780                         return;
 781                 }
 782                 __sk_nulls_del_node_init_rcu(sk);
 783                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 784                 spin_unlock_bh(lock);
 785         }
 786 }
 787 EXPORT_SYMBOL_GPL(inet_unhash);
 788
 789 static bool inet_bind2_bucket_match(const struct inet_bind2_bucket *tb,
 790                                     const struct net *net, unsigned short port,
 791                                     int l3mdev, const struct sock *sk)
 792 {
 793 #if IS_ENABLED(CONFIG_IPV6)
 794         if (sk->sk_family == AF_INET6)
 795                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 796                         tb->l3mdev == l3mdev &&
 797                         ipv6_addr_equal(&tb->v6_rcv_saddr, &sk->sk_v6_rcv_saddr);
 798         else
 799 #endif
 800                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 801                         tb->l3mdev == l3mdev && tb->rcv_saddr == sk->sk_rcv_saddr;
 802 }
 803
 804 bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const struct net *net,
 805                                       unsigned short port, int l3mdev, const struct sock *sk)
 806 {
 807 #if IS_ENABLED(CONFIG_IPV6)
 808         struct in6_addr addr_any = {};
 809
 810         if (sk->sk_family == AF_INET6)
 811                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 812                         tb->l3mdev == l3mdev &&
 813                         ipv6_addr_equal(&tb->v6_rcv_saddr, &addr_any);
 814         else
 815 #endif
 816                 return net_eq(ib2_net(tb), net) && tb->port == port &&
 817                         tb->l3mdev == l3mdev && tb->rcv_saddr == 0;
 818 }
 819
 820 /* The socket's bhash2 hashbucket spinlock must be held when this is called */
 821 struct inet_bind2_bucket *
 822 inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct net *net,
 823                        unsigned short port, int l3mdev, const struct sock *sk)
 824 {
 825         struct inet_bind2_bucket *bhash2 = NULL;
 826
 827         inet_bind_bucket_for_each(bhash2, &head->chain)
 828                 if (inet_bind2_bucket_match(bhash2, net, port, l3mdev, sk))
 829                         break;
 830
 831         return bhash2;
 832 }
 833
 834 struct inet_bind_hashbucket *
 835 inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, int port)
 836 {
 837         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 838         u32 hash;
 839 #if IS_ENABLED(CONFIG_IPV6)
 840         struct in6_addr addr_any = {};
 841
 842         if (sk->sk_family == AF_INET6)
 843                 hash = ipv6_portaddr_hash(net, &addr_any, port);
 844         else
 845 #endif
 846                 hash = ipv4_portaddr_hash(net, 0, port);
 847
 848         return &hinfo->bhash2[hash & (hinfo->bhash_size - 1)];
 849 }
 850
 851 int inet_bhash2_update_saddr(struct inet_bind_hashbucket *prev_saddr, struct sock *sk)
 852 {
 853         struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);
 854         struct inet_bind2_bucket *tb2, *new_tb2;
 855         int l3mdev = inet_sk_bound_l3mdev(sk);
 856         struct inet_bind_hashbucket *head2;
 857         int port = inet_sk(sk)->inet_num;
 858         struct net *net = sock_net(sk);
 859
 860         /* Allocate a bind2 bucket ahead of time to avoid permanently putting
 861          * the bhash2 table in an inconsistent state if a new tb2 bucket
 862          * allocation fails.
 863          */
 864         new_tb2 = kmem_cache_alloc(hinfo->bind2_bucket_cachep, GFP_ATOMIC);
 865         if (!new_tb2)
 866                 return -ENOMEM;
 867
 868         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
 869
 870         if (prev_saddr) {
 871                 spin_lock_bh(&prev_saddr->lock);
 872                 __sk_del_bind2_node(sk);
 873                 inet_bind2_bucket_destroy(hinfo->bind2_bucket_cachep,
 874                                           inet_csk(sk)->icsk_bind2_hash);
 875                 spin_unlock_bh(&prev_saddr->lock);
 876         }
 877
 878         spin_lock_bh(&head2->lock);
 879         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
 880         if (!tb2) {
 881                 tb2 = new_tb2;
 882                 inet_bind2_bucket_init(tb2, net, head2, port, l3mdev, sk);
 883         }
 884         sk_add_bind2_node(sk, &tb2->owners);
 885         inet_csk(sk)->icsk_bind2_hash = tb2;
 886         spin_unlock_bh(&head2->lock);
 887
 888         if (tb2 != new_tb2)
 889                 kmem_cache_free(hinfo->bind2_bucket_cachep, new_tb2);
 890
 891         return 0;
 892 }
 893 EXPORT_SYMBOL_GPL(inet_bhash2_update_saddr);
 894
 895 /* RFC 6056 3.3.4.  Algorithm 4: Double-Hash Port Selection Algorithm
 896  * Note that we use 32bit integers (vs RFC 'short integers')
 897  * because 2^16 is not a multiple of num_ephemeral and this
 898  * property might be used by clever attacker.
 899  * RFC claims using TABLE_LENGTH=10 buckets gives an improvement, though
 900  * attacks were since demonstrated, thus we use 65536 instead to really
 901  * give more isolation and privacy, at the expense of 256kB of kernel
 902  * memory.
 903  */
 904 #define INET_TABLE_PERTURB_SHIFT 16
 905 #define INET_TABLE_PERTURB_SIZE (1 << INET_TABLE_PERTURB_SHIFT)
 906 static u32 *table_perturb;
 907
 908 int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 909                 struct sock *sk, u64 port_offset,
 910                 int (*check_established)(struct inet_timewait_death_row *,
 911                         struct sock *, __u16, struct inet_timewait_sock **))
 912 {
 913         struct inet_hashinfo *hinfo = death_row->hashinfo;
 914         struct inet_bind_hashbucket *head, *head2;
 915         struct inet_timewait_sock *tw = NULL;
 916         int port = inet_sk(sk)->inet_num;
 917         struct net *net = sock_net(sk);
 918         struct inet_bind2_bucket *tb2;
 919         struct inet_bind_bucket *tb;
 920         bool tb_created = false;
 921         u32 remaining, offset;
 922         int ret, i, low, high;
 923         int l3mdev;
 924         u32 index;
 925
 926         if (port) {
 927                 head = &hinfo->bhash[inet_bhashfn(net, port,
 928                                                   hinfo->bhash_size)];
 929                 tb = inet_csk(sk)->icsk_bind_hash;
 930                 spin_lock_bh(&head->lock);
 931                 if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
 932                         inet_ehash_nolisten(sk, NULL, NULL);
 933                         spin_unlock_bh(&head->lock);
 934                         return 0;
 935                 }
 936                 spin_unlock(&head->lock);
 937                 /* No definite answer... Walk to established hash table */
 938                 ret = check_established(death_row, sk, port, NULL);
 939                 local_bh_enable();
 940                 return ret;
 941         }
 942
 943         l3mdev = inet_sk_bound_l3mdev(sk);
 944
 945         inet_get_local_port_range(net, &low, &high);
 946         high++; /* [32768, 60999] -> [32768, 61000[ */
 947         remaining = high - low;
 948         if (likely(remaining > 1))
 949                 remaining &= ~1U;
 950
 951         net_get_random_once(table_perturb,
 952                             INET_TABLE_PERTURB_SIZE * sizeof(*table_perturb));
 953         index = port_offset & (INET_TABLE_PERTURB_SIZE - 1);
 954
 955         offset = READ_ONCE(table_perturb[index]) + (port_offset >> 32);
 956         offset %= remaining;
 957
 958         /* In first pass we try ports of @low parity.
 959          * inet_csk_get_port() does the opposite choice.
 960          */
 961         offset &= ~1U;
 962 other_parity_scan:
 963         port = low + offset;
 964         for (i = 0; i < remaining; i += 2, port += 2) {
 965                 if (unlikely(port >= high))
 966                         port -= remaining;
 967                 if (inet_is_local_reserved_port(net, port))
 968                         continue;
 969                 head = &hinfo->bhash[inet_bhashfn(net, port,
 970                                                   hinfo->bhash_size)];
 971                 spin_lock_bh(&head->lock);
 972
 973                 /* Does not bother with rcv_saddr checks, because
 974                  * the established check is already unique enough.
 975                  */
 976                 inet_bind_bucket_for_each(tb, &head->chain) {
 977                         if (inet_bind_bucket_match(tb, net, port, l3mdev)) {
 978                                 if (tb->fastreuse >= 0 ||
 979                                     tb->fastreuseport >= 0)
 980                                         goto next_port;
 981                                 WARN_ON(hlist_empty(&tb->owners));
 982                                 if (!check_established(death_row, sk,
 983                                                        port, &tw))
 984                                         goto ok;
 985                                 goto next_port;
 986                         }
 987                 }
 988
 989                 tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
 990                                              net, head, port, l3mdev);
 991                 if (!tb) {
 992                         spin_unlock_bh(&head->lock);
 993                         return -ENOMEM;
 994                 }
 995                 tb_created = true;
 996                 tb->fastreuse = -1;
 997                 tb->fastreuseport = -1;
 998                 goto ok;
 999 next_port:
1000                 spin_unlock_bh(&head->lock);
1001                 cond_resched();
1002         }
1003
1004         offset++;
1005         if ((offset & 1) && remaining > 1)
1006                 goto other_parity_scan;
1007
1008         return -EADDRNOTAVAIL;
1009
1010 ok:
1011         /* Find the corresponding tb2 bucket since we need to
1012          * add the socket to the bhash2 table as well
1013          */
1014         head2 = inet_bhashfn_portaddr(hinfo, sk, net, port);
1015         spin_lock(&head2->lock);
1016
1017         tb2 = inet_bind2_bucket_find(head2, net, port, l3mdev, sk);
1018         if (!tb2) {
1019                 tb2 = inet_bind2_bucket_create(hinfo->bind2_bucket_cachep, net,
1020                                                head2, port, l3mdev, sk);
1021                 if (!tb2)
1022                         goto error;
1023         }
1024
1025         /* Here we want to add a little bit of randomness to the next source
1026          * port that will be chosen. We use a max() with a random here so that
1027          * on low contention the randomness is maximal and on high contention
1028          * it may be inexistent.
1029          */
1030         i = max_t(int, i, (prandom_u32() & 7) * 2);
1031         WRITE_ONCE(table_perturb[index], READ_ONCE(table_perturb[index]) + i + 2);
1032
1033         /* Head lock still held and bh's disabled */
1034         inet_bind_hash(sk, tb, tb2, port);
1035
1036         spin_unlock(&head2->lock);
1037
1038         if (sk_unhashed(sk)) {
1039                 inet_sk(sk)->inet_sport = htons(port);
1040                 inet_ehash_nolisten(sk, (struct sock *)tw, NULL);
1041         }
1042         if (tw)
1043                 inet_twsk_bind_unhash(tw, hinfo);
1044         spin_unlock(&head->lock);
1045         if (tw)
1046                 inet_twsk_deschedule_put(tw);
1047         local_bh_enable();
1048         return 0;
1049
1050 error:
1051         spin_unlock(&head2->lock);
1052         if (tb_created)
1053                 inet_bind_bucket_destroy(hinfo->bind_bucket_cachep, tb);
1054         spin_unlock_bh(&head->lock);
1055         return -ENOMEM;
1056 }
1057
1058 /*
1059  * Bind a port for a connect operation and hash it.
1060  */
1061 int inet_hash_connect(struct inet_timewait_death_row *death_row,
1062                       struct sock *sk)
1063 {
1064         u64 port_offset = 0;
1065
1066         if (!inet_sk(sk)->inet_num)
1067                 port_offset = inet_sk_port_offset(sk);
1068         return __inet_hash_connect(death_row, sk, port_offset,
1069                                    __inet_check_established);
1070 }
1071 EXPORT_SYMBOL_GPL(inet_hash_connect);
1072
1073 static void init_hashinfo_lhash2(struct inet_hashinfo *h)
1074 {
1075         int i;
1076
1077         for (i = 0; i <= h->lhash2_mask; i++) {
1078                 spin_lock_init(&h->lhash2[i].lock);
1079                 INIT_HLIST_NULLS_HEAD(&h->lhash2[i].nulls_head,
1080                                       i + LISTENING_NULLS_BASE);
1081         }
1082 }
1083
1084 void __init inet_hashinfo2_init(struct inet_hashinfo *h, const char *name,
1085                                 unsigned long numentries, int scale,
1086                                 unsigned long low_limit,
1087                                 unsigned long high_limit)
1088 {
1089         h->lhash2 = alloc_large_system_hash(name,
1090                                             sizeof(*h->lhash2),
1091                                             numentries,
1092                                             scale,
1093                                             0,
1094                                             NULL,
1095                                             &h->lhash2_mask,
1096                                             low_limit,
1097                                             high_limit);
1098         init_hashinfo_lhash2(h);
1099
1100         /* this one is used for source ports of outgoing connections */
1101         table_perturb = alloc_large_system_hash("Table-perturb",
1102                                                 sizeof(*table_perturb),
1103                                                 INET_TABLE_PERTURB_SIZE,
1104                                                 0, 0, NULL, NULL,
1105                                                 INET_TABLE_PERTURB_SIZE,
1106                                                 INET_TABLE_PERTURB_SIZE);
1107 }
1108
1109 int inet_hashinfo2_init_mod(struct inet_hashinfo *h)
1110 {
1111         h->lhash2 = kmalloc_array(INET_LHTABLE_SIZE, sizeof(*h->lhash2), GFP_KERNEL);
1112         if (!h->lhash2)
1113                 return -ENOMEM;
1114
1115         h->lhash2_mask = INET_LHTABLE_SIZE - 1;
1116         /* INET_LHTABLE_SIZE must be a power of 2 */
1117         BUG_ON(INET_LHTABLE_SIZE & h->lhash2_mask);
1118
1119         init_hashinfo_lhash2(h);
1120         return 0;
1121 }
1122 EXPORT_SYMBOL_GPL(inet_hashinfo2_init_mod);
1123
1124 int inet_ehash_locks_alloc(struct inet_hashinfo *hashinfo)
1125 {
1126         unsigned int locksz = sizeof(spinlock_t);
1127         unsigned int i, nblocks = 1;
1128
1129         if (locksz != 0) {
1130                 /* allocate 2 cache lines or at least one spinlock per cpu */
1131                 nblocks = max(2U * L1_CACHE_BYTES / locksz, 1U);
1132                 nblocks = roundup_pow_of_two(nblocks * num_possible_cpus());
1133
1134                 /* no more locks than number of hash buckets */
1135                 nblocks = min(nblocks, hashinfo->ehash_mask + 1);
1136
1137                 hashinfo->ehash_locks = kvmalloc_array(nblocks, locksz, GFP_KERNEL);
1138                 if (!hashinfo->ehash_locks)
1139                         return -ENOMEM;
1140
1141                 for (i = 0; i < nblocks; i++)
1142                         spin_lock_init(&hashinfo->ehash_locks[i]);
1143         }
1144         hashinfo->ehash_locks_mask = nblocks - 1;
1145         return 0;
1146 }
1147 EXPORT_SYMBOL_GPL(inet_ehash_locks_alloc);